sensors

Article
A Novel Chiller Sensors Fault Diagnosis Method
Based on Virtual Sensors
Long Gao 1, * , Donghui Li 1 , Ding Li 1 , Lele Yao 1 , Limei Liang 2 and Yanan Gao 1
1 School of Electrical and Information Engineering, Tianjin University, Nankai District, Tianjin 300072, China
2 School of Control Science and Engineering, Shandong University, Lixia District, Jinan 250061,
Shandong, China
* Correspondence: gao_long@tju.edu.cn; Tel.: +86-188-2203-1612




Received: 12 June 2019; Accepted: 3 July 2019; Published: 8 July 2019


Abstract: Sensor fault detection and diagnosis (FDD) has great significance for ensuring the energy
saving and normal operation of the air conditioning system. Chiller systems serving as an important
part of central air conditioning systems are the major energy consumer in commercial and industrial
buildings. In order to ensure the normal operation of the chiller system, virtual sensors have been
proposed to detect and diagnose sensor faults. However, the performance of virtual sensors could
be easily impacted by abnormal data. To solve this problem, virtual sensors combined with the
maximal information coefficient (MIC) and a long short-term memory (LSTM) network is proposed
for chiller sensor fault diagnosis. Firstly, MIC, which has the ability to quantify the degree of
relevance in a data set, is applied to examine all potentially interesting relationships between
sensors. Subsequently, sensors with high correlation are divided into several groups by the grouping
thresholds. Two virtual sensors, which are constructed in each group by LSTM with different input
sensors and corresponding to the same physical sensor, could have the ability to predict the value of
physical sensors. High correlation sensors in each group improve the fitting effect of virtual sensors.
Finally, sensor faults can be diagnosed by the absolute deviation which is generated by comparing the
virtual sensors’ output with the actual value measured from the air-cooled chiller. The performance
of the proposed method is evaluated by using a real data set. Experimental results indicate that
virtual sensors can be well constructed and the proposed method achieves a significant performance
along with a low false alarm rate.

Keywords: virtual sensors; fault detection and diagnosis (FDD); maximum information
coefficient (MIC); low false alarm rate; air-cooled chiller

1. Introduction
Poorly maintained and improperly controlled equipment wastes an estimated 15% to 30%
of energy used in commercial buildings [1]. Heating, ventilation, and air conditioning (HVAC),
which maintain comfortable and healthy indoor thermal environments, is an important part of public
and private buildings [2]. As an important part of central air conditioning systems, the chiller is the
major energy consumer [3]. Hence, it is of vital importance to have a fault detection and diagnosis
(FDD) method to maintain optimal operation for chiller systems.
FDD provides a cornerstone for the condition-based maintenance of engineered systems and
has been an active area of research and development in the aerospace, process controls, automotive,
manufacturing, nuclear, and national defense fields [4–8]. There are many methods to realize FDD in
different fields. Data driven methods, rule based methods and model based methods have been used
in building systems [9–11].

Sensors 2019, 19, 3013; doi:10.3390/s19133013 www.mdpi.com/journal/sensors

Sensors 2019, 19, 3013 2 of 15

With the rapid development of the computer’s calculation ability, data driven methods have been
widely developed in recent years. Han et al. present a novel FDD strategy, which combines the principle
component analysis (PCA) feature extraction technology and the multiclass support vector machine
(SVM) classification algorithm for vapor-compression refrigeration systems [12]. In order to fully
capture the data characteristics, Li proposed a novel data-temporal attention network based strategy
for the fault diagnosis of chiller sensors [13]. Fan et al. present a back-propagation neural network
black box model using wavelet analysis and fuzzy logic to detect and diagnose faults in an air handling
unit [14]. In Reference [15], a dynamic model combined with a data driven method is used to estimate
the remaining useful life which does not require prior knowledge of the degradation phenomena.
In another respect, hardware redundancy was first developed to diagnose faults. It has high
reliability and can directly isolate faults [16]. Regrettably, hardware redundancy leads to costly and
time-consuming processes [17]. To solve this problem, virtual sensors have been proposed for FDD [18]
and applied to many fields [19–22]. Virtual sensors, which are able to estimate various phenomena
that are difficult or expensive to measure, can be mass-produced by using black box or gray box
models along with other existing physical sensors in building systems [23]. Li et al., review virtual
sensing techniques and early applications for buildings [24]. In building systems, virtual sensors can
be constructed with historical data which are stored by a monitoring system. The method proposed
in Reference [18] is applied to detect and diagnose faults in ventilation units using virtual sensors.
The physical sensors can be predicted by virtual sensors with satisfactory accuracy. The method
proposed in Reference [25] is applied to exploit physical relations inside the unit using linear regression
virtual sensors. Vasso et al. present a local virtual sensor agent for diagnosing sensor faults in HVAC
systems, and compensating for their effects on the distributed control architecture [26]. Kusiak et
al. have constructed sensor models for predicting temperature, CO2 , and relative humidity by data
mining algorithms. It can be applied to HVAC systems in various buildings [27]. However, all of the
above virtual sensors’ performances could be easily impacted by input parameters, such as historical
data of a building system and external factors [23].
In this paper, a novel chiller sensor fault diagnosis method based on a virtual sensor technique
and a data driven method is proposed. Firstly, a large amount of sensor data can be stored when the
chiller system is operating. According to the data collected from the chiller system, sensors can be
divided into several groups by the maximal information coefficient (MIC) and the sensor grouping
threshold. MIC is used to explore relationships among different sensors, which can better describe
the nonlinear relationship in data set than traditional methods such as Pearson correlation. Sensors
in each group have high correlation which is helpful for improving the performance of the virtual
sensors. Subsequently, virtual sensors are constructed by a long short-term memory network (LSTM)
in each group. Deep learning models are capable of automatic and deep mining feature information,
which have made achievements in many fields [28–30]. As a deep learning model, LSTM could be
capable of automatic and deep mining feature information in sensor data, which can better improve
the performance of virtual sensors. In the end, the trained networks could properly fit the chiller
system. The fault can be diagnosed by the absolute deviation which is generated by comparing the
predicted output of the virtual sensors with the actual value.
The contributions of the paper are as follows:

1. In order to ensure the performance of virtual sensors, MIC is used to examine potentially
interesting relationships between sensors. Chiller sensors with high MIC scores are divided
into the same groups. This could dramatically improve the fitting effect of virtual sensors by
constructing them in the same group.
2. The performance of virtual sensors could be easily impacted by the input sensors. In order to
reduce the false alarm rate, two virtual sensors that have different input sensors are constructed
for the same physical sensor. When the two deviations between the corresponding physical
sensor and the two virtual sensors both exceed the thresholds, the physical sensor is considered
as a fault state.
Sensors 2019, 19, 3013 3 of 15

3. The LSTM model, which can better extract discriminating features from the sensor data, is used
to construct the virtual sensors. It could further improve the fitting effect of virtual sensors.
The rest of this paper is organized as follows: Section 2 gives the description and coupling
characteristic analysis of the air-cooled chiller system; the proposed chiller sensor fault diagnosis
method is introduced in Section 3; Section 4 verifies the effectiveness of the proposed method
on the sensor fault diagnosis performance along with low false alarm rate; and Section 5 draws
the conclusions.

2. Coupling Characteristic Analysis of the Air-Cooled Chiller System

Figure 1 shows the schematic of an existing air-cooled chiller system. The main equipment of the
air-cooled chiller includes a scroll compressor, air-cooled condenser, throttle device and evaporator.
The throttling device is provided with three different throttling modes: electronic expansion valve,
thermal expansion valve and electric needle throttle. This paper mainly uses an electronic expansion
valve to throttle the refrigerant. The chilled water, which is powered by the pump, circulates
by the green line. When the water temperature in the water tank is lowered, the electric heater,
which maintains the temperature of the water tank, is used to compensate the heat consumption.

Air-cooled 3 Dry filter 10 4

Thermal expansion valve

Electronic expansion valve

fin condenser Electric needle
throttle valve

Liquid storage
tank

9
11 5
2

evaporator
Water tank
Scroll
compressor
1 8
6 Electric heating
pump 7

Figure 1. Schematic diagram of the air-cooled chiller system. The system consists of refrigerant closed
loop and chilled water closed loop.

The heat is transferred from the low-temperature heat source to the high-temperature heat source
via refrigerant circulation loop. In the circulation loop of the refrigerant, the energy conservation can
be written as
Qe + Qcom = Qc , (1)

where Qcom represents the electric power consumed by the compressor, Qe represents the amount of
cold generated in the evaporator and Qc represents the heat released by the condenser.
Sensors 2019, 19, 3013 4 of 15

The refrigerant circulates via the compressor, condenser, expansion valve and evaporator while
performing energy transfer with the chilled water. The pressure, enthalpy and other parameters of the
refrigerant in the condenser can be affected by the output parameters of the compressor, the outlet
refrigerant enthalpy, the mass flow rate, and so forth. Conversely, the compressor power and the
refrigerant mass flow can be changed by the compressor input parameters such as the condenser
condensing pressure. The parameters, such as the electronic expansion valve mass flow and the outlet
enthalpy, can be affected by the output parameters of the condenser, the condensing pressure, the outlet
refrigerant enthalpy, and so forth. The condensing pressure can also be affected by the electronic
expansion valve mass flow. The electronic expansion valve output parameters and the enthalpy affect
the evaporation pressure, the outlet enthalpy and other parameters. The evaporation pressure as the
expansion valve input can also affect the electronic expansion valve mass flow. There is also a coupling
relationship between the compressor and the evaporator through the evaporation pressure and the
mass flow of the compressor. Thus, there is a complicated coupling relationship between the various
equipment of the chiller system.

3. Methods

3.1. Maximal Information Coefficient

Due to the complex coupling characteristic of the air-cooled chiller system, MIC which has the
ability to examine all potentially interesting relationships is used to explore relationships among
different sensors. The calculation of MIC is based on mutual information (MI). MI tells us how much
knowing one variable reduces our uncertainty about the other [31]. And the mutual information of
two variables X and Y can be defined as

u( x, y)
Z Z
I ( X, Y ) = u( x, y)log dxdy, (2)
u x ( x )uy (y)
R
where u( x, y) is the joint probability density function of X and Y, u x ( x ) = u( x, y) dy and
R
uy (y) = u( x, y) dx.
In order to get the MIC score, MI is first achieved by exploring all the grids by using different
partition schemes, then the MIC is normalized between 0 and 1 by dividing log2 (min( X, Y )), and the
maximum is chosen as the MIC score. Thus the MIC score can be calculated as [32]

I ( D, X, Y )
M( D) = max , (3)
XY < B(n) log2 (min( X, Y ))

where B is a function of sample size n, usually B = n0.6 ; I ( D, X, Y ) is the maximum mutual information
value that falls into the mesh region D. Variables X and Y are independent of each other when MIC is
equal to 0. And some kinds of functional relationship are existed between X and Y variables when
MIC is equal to 1.
Raw operating data, which include eleven sensors, compressor frequency Cs and electron
expansion valve Eev , are provided by the air-cooled chiller system. In this system, temperature
is measured using PT-100, whose average error is around 0.3 ◦ C. The pressure on the scroll compressor
and throttle device are measured by pressure transducers with the accuracy 0.5%. The descriptions of
the eleven sensors have been listed in Table 1.
Sensors 2019, 19, 3013 5 of 15

Table 1. The descriptions of the eleven sensors.

No. Sensors Descriptions Unit

1 Tsuc Compressor suction temperature ◦C
2 Tdis Compressor discharge temperature ◦C
3 Tcon_out Condenser-air temperature at the outlet ◦C
4 Tre f _b Refrigerant temperature before throttling ◦C

5 Tre f _a Refrigerant temperature after throttling ◦C

6 Tchw_s Chilled-water supply temperature ◦C

7 Tchw_r Chilled-water return temperature ◦C
8 Psuc Compressor suction pressure MPa
9 Pdis Compressor discharge pressure MPa
10 Pth_in Inlet pressure of the throttle device MPa
11 Pth_out Outlet pressure of the throttle device MPa

The maximum mutual information is calculated between any sensors in Table 2. Multiple chiller
sensors are divided in the same group to ensure the correlation between the sensors, which the MIC score
is greater than grouping threshold. For instance, sensors can be divided into {Tsuc , Tchw_s , Tchw_r , Pth_out }
and {Psuc , Tcon_out , Tre f _b , Pdis } for virtual senors Tsuc and Psuc , when the grouping threshold takes 0.8.
In order to verify MIC mining relationship capabilities between the sensors, density plots of Tsuc and Tdis ,
Tdis and Tcon_out are shown in Figure 2. As we can see, Tsuc and Tdis which the MIC score is equal to 0.239
are independent on each other, while Tdis and Tcon_out which the MIC score is equal to 0.812 are obviously
highly correlated.

Table 2. MIC of different sensors.

Tsuc Tdis Tcon_out Tre f _b Tre f _a Tchw_s Tchw_r Psuc Pdis Pth_in Pth_out
Tsuc 1 0.239 0.264 0.279 0.432 0.850 0.881 0.267 0.271 0.618 0.832
Tdis 0.239 1 0.812 0.652 0.307 0.237 0.237 0.750 0.827 0.319 0.092
Tcon_out 0.264 0.812 1 0.836 0.249 0.305 0.299 0.922 0.898 0.255 0.160
Tre f _b 0.279 0.652 0.836 1 0.289 0.337 0.328 0.864 0.906 0.273 0.188
Tre f _a 0.432 0.307 0.249 0.289 1 0.348 0.422 0.255 0.263 0.906 0.805
Tchw_s 0.850 0.237 0.305 0.337 0.348 1 0.886 0.334 0.333 0.335 0.516
Tchw_r 0.881 0.237 0.299 0.328 0.422 0.886 1 0.325 0.328 0.431 0.499
Psuc 0.267 0.750 0.922 0.864 0.255 0.334 0.325 1 0.940 0.260 0.180
Pdis 0.271 0.827 0.898 0.906 0.263 0.333 0.328 0.940 1 0.262 0.183
Pth_in 0.618 0.319 0.255 0.273 0.906 0.335 0.431 0.260 0.262 1 0.135
Pth_out 0.832 0.092 0.160 0.188 0.805 0.516 0.499 0.180 0.183 0.135 1
Condenser-air temperature
Compressor discharge

at the outlet (°С)

temperature (°С)

Compressor suction temperature (°С) Compressor discharge temperature (°С)

Figure 2. Density plots of Tsuc and Tdis , Tdis and Tcon_out .
Sensors 2019, 19, 3013 6 of 15

3.2. Virtual Sensors

The LSTM model is used to construct the virtual sensors for eleven physical sensors. It can better
extract feature information among different sensors in the chiller system. As shown in Figure 3, the key
to LSTM is the LSTM cell which can decide whether to maintain state information from the prior step.
LSTM cells contain input gate, forget gate and output gate. The input gate is used to decide the input
information to save the state of unit. The forget gate is used to decide the state of unit at the last time.
The output gate is used to decide the output of LSTM cell. Outputs ct and ht are recurrently connected
to the inputs of block. The LSTM cell with forward propagation is calculated as follows:

f t = σ (W f ∗ [ h t − 1 , x t ] + b f ) , (4)

it = σ (Wi ∗ [ht−1 , xt ] + bi ), (5)

ot = σ (Wo ∗ [ht−1 , xt ] + bo ), (6)

c t = f t ∗ c t −1 + i t ∗ g t , (7)

gt = tanh(Wc ∗ [ht−1 , xt ] + bc ), (8)

ht = ot ∗ tanh(ct ). (9)

Figure 3. LSTM cell.

The Formulas (4)–(6) are forget gate, input gate and output gate respectively. W f , Wi and Wc are
the weights of forget gate, input gate and output gate respectively. σ (·) and tanh(·) are the sigmoid
function and hyperbolic tangent respectively.
Herein, the input of the LSTM model includes the sensor historical data in same group, compressor
frequency and electron expansion valve. And the output of the LSTM model constructs corresponding
virtual sensors. Two virtual sensors, which contain different input sensors and correspond to the same
physical sensor, are constructed in the same group. The maximum correlation ensures the performance of
the two virtual sensors with different input physical sensors. For instance, if the grouping threshold takes
0.3 in Table 3, the chiller sensors can be divided into two groups: {Tsuc , Tre f _a , Tchw_s , Tchw_r , Pth_in , Pth_out }
0 and T 00 for T
and {Psuc , Tdis , Tcon_out , Tre f _b , Tchw_s , Tchw_r , Pdis }. Virtual sensor Tsuc suc suc are constructed as

−−→
VT0 suc = f (VT1suc ), (10)
−−→
VT00suc = f (VT2suc ), (11)
−−→
VT1suc = ( Tre f _a , Tchw_s , Tchw_r , Cs , Eev ), (12)
−−→
VT2suc = ( Pth_in , Pth_out , Cs , Eev ), (13)
Sensors 2019, 19, 3013 7 of 15

−−→ −−→
where VT1suc and VT2suc are acquired sequentially from the corresponding groups and represents the
sensors historical data. VT0 suc and VT00suc represent corresponding virtual sensors and f (·) represents
LSTM model. The actual value and predictive value for Tsuc are shown in Figure 4.

Table 3. Different groups for Tsuc and Psuc .

Grouping Threshold Group

{Tsuc , Tchw_s , Tchw_r , Pth_out }
0.8
{ Psuc , Tcon_out , Tre f _b , Pdis }
{Tsuc , Tchw_s , Tchw_r , Pth_in , Pth_out }
0.6
{ Psuc , Tdis , Tcon_out , Tre f _b , Pdis }
{Tsuc , Tre f _a , Tchw_s , Tchw_r , Pth_in , Pth_out }
0.3
{ Psuc , Tdis , Tcon_out , Tre f _b , Tchw_s , Tchw_r , Pdis }

Physical sensor Physical sensor

Virtual sensor Virtual sensor

(a) (b)
Figure 4. (a) Predictive result of virtual sensor VT1suc . (b) Predictive result of virtual sensor VT2suc .

3.3. The Threshold and Procedure of Fault Diagnosis

In the actual implementation, the fault can be diagnosed by the absolute deviation between the
virtual sensors and the physical sensors. A fault diagnosis threshold is calculated to automatically
diagnose sensor fault. The fault diagnosis threshold of the l-th sensor is calculated as

l
VFD = max( abs(vlp − alp )), (14)
p∈ST

where vlp is the predictive value of virtual sensors, alp is the actual value of chiller sensors and ST
is a set to determinate fault diagnosis threshold. Different group thresholds can be obtained for
different sensors.
Figure 5 shows the flow chart of fault diagnosis. The proposed fault diagnosis method for chiller
system includes three steps:

Step 1: A physical sensor is selected from eleven sensors. Two virtual sensors, which have been
constructed in the training period, are used to predict the value of this physical sensor.
Step 2: Deviations between virtual and physical sensors are calculated and compared with the
fault diagnosis threshold.
Step 3: Obviously, the physical sensor is not considered as a fault state when no deviations exceed
the threshold. On the contrary, a sensor fault occurs when the deviations both exceed the
fault diagnosis threshold. Input sensors are considered as a fault state if only one absolute
deviation exceeds the threshold, because two virtual sensors have different input sensors.
Under this situation, another physical sensor from input sensors is selected and step 2 will
be repeated to predict the value of another physical sensor.
Sensors 2019, 19, 3013 8 of 15

Start

Step 1
Select a physical sensor

Step 3
Two virtual sensors Choose another physical sensor
of this physical sensor

Another sensors fault

Step 2 Computational deviations and

comparison with threshold
Y

The deviations Only one

exceed the threshold? N
exceeds the threshold?

Y N

Step 3
Sensor fault No fault

Figure 5. Flow chart of fault diagnosis.

4. Results and Discussion

4.1. Experimental Data

The experimental data are acquired from the device in the Honeywell Home and Building Control
Laboratory in Tianjin University, China. During the normal operation of the device, the chiller system
can store thirteen kinds of measurement data such as different sensor values and electronic expansion
valve opening degree which are highly correlated with the sensor reading. The experimental data,
which obtain 45,000 sets of time series, are collected at 2 min intervals and divided into training sets,
fault diagnosis threshold determination sets and test sets.
Before the experiment began, the raw data need to be standardized and filter out samples
with poor quality, such as those during chiller start up and shutdown time periods. Based on the
measurement accuracy of temperatures and pressure sensors, eleven sensor faults are introduced by
adding the fixed bias in this paper. Various magnitudes of biases, which are determined by the reading
range of each sensor, have been added to each sensor in Table 4.

Table 4. Biases added in different sensors.

Sensors Unit Biases

Tsuc ◦C ±0.5, ±0.65, ±0.8, ±0.95
Tdis ◦C ±1.5, ±1.9, ±2.3, ±2.7
Tcon_out ◦C ±1.0, ±1.3, ±1.6, ±1.9
Tre f _b ◦C ±1.0, ±1.3, ±1.6, ±1.9
Tre f _a ◦C ±0.4, ±0.5, ±0.6, ±0.7
Tchw_s ◦C ±0.4, ±0.5, ±0.6, ±0.7
Tchw_r ◦C ±0.4, ±0.5, ±0.6, ±0.7
Psuc MPa ±0.035, ±0.04, ±0.045, ±0.05
Pdis MPa ±0.07, ±0.08, ±0.09, ±0.10
Pth_in MPa ±0.07, ±0.08, ±0.09, ±0.10
Pth_out MPa ±0.035, ±0.04, ±0.045, ±0.05
Sensors 2019, 19, 3013 9 of 15

4.2. Performance Comparison

To verify the low false alarm rate and fault diagnosis performance of this method, Tsuc and Psuc are
chosen as instances. As described in Section 3.1, different groups for Tsuc and Psuc have been obtained
and listed in Table 3 when different grouping thresholds are used. As can be seen, the number of the
sensors has decreased along with the increase of the grouping thresholds. A high correlation between
sensors will contribute to the constructing of a virtual sensor in the same group. All chiller sensors are
divided into the same group, if the grouping threshold takes 0.

4.2.1. Verification of Low False Alarm Rate

We assume that only one sensor is considered as a fault state. Input sensors Tchw_s and Tcon_out
with minimum biases are used to verify the low false alarm rate of this method. Sensor grouping
threshold takes 0.8. In the figure below, the first half time series are normal samples and the last half
are fault samples. Figure 6 shows the absolute deviation of Tsuc and Psuc when the corresponding input
sensors Tchw_s and Tcon_out are faulty. As can be seen in Figure 6a, if only one virtual sensor is used,
the absolute deviation will increases after fault occurs. It means that the fault diagnosis performance
will be affected by the input sensor. However, another sensor fault occurs if only one absolute deviation
exceeds the fault diagnosis threshold, as depicted in Figures 5 and 6b,c. The false alarm rate of Tsuc is
reduced from 28.0% to 0.0%. In the same way, the false alarm rate of Psuc is reduced from 41.0% to 0.0%.

0.6 0.6
The absolute deviation

The absolute deviation

The absolute deviation

0.6

0.4 0.4
0.4

0.2 0.2 0.2

0.0 0.0 0.0

Time series Time series Time series
(a) (b) (c)
0.03 0.03
The absolute deviation

The absolute deviation

The absolute deviation

0.03

0.02 0.02
0.02

0.01 0.01 0.01

0.00 0.00 0.00

Time series Time series Time series

(d) (e) (f)

Figure 6. The absolute deviation of Tsuc and Psuc with different virtual sensors. (a) The absolute
deviation between the physical sensor Tsuc and the only virtual sensor. (b) The absolute deviation
between Tsuc and the first of the two virtual sensor. (c) The absolute deviation between Tsuc and the
second of the two virtual sensor after grouping. (d) The absolute deviation between the physical sensor
Psuc and the only virtual sensor. (e) The absolute deviation between Psuc and the first of the two virtual
sensor. (f) The absolute deviation between Psuc and the second of the two virtual sensor.

To further verify the performance of this method, the false alarm rate of all chiller sensors is listed
in Table 5. As we can see, the false alarm rate is close to 0 when two virtual sensors with different
input sensor are used. For instance, the false alarm rate of Tcon_out will reduce from 18.0% to 3.0%,
if two virtual sensors for Tcon_out are used. The high false alarm rate will occur if only one virtual
sensor is used. Therefore, two virtual sensors with different input sensors can achieve a low false
alarm rate.
Sensors 2019, 19, 3013 10 of 15

Table 5. False alarm rate of chiller sensors.

Tdis Tcon_out Tre f _b Tre f _a Tchw_s Tchw_r Pdis Pth_in Pth_out

Only one virtual sensor 35.0% 18.0% 26.0% 52.0% 44.0% 27.0% 39.0% 11.0% 23.0%
Two virtual sensors 0.00% 3.0% 0.00% 0.00% 0.00% 2.0% 4.0% 0.00% 0.00%

4.2.2. Fault Diagnosis Performance

As a deep learning model, the LSTM model, which is developed to address sequential data with
its ability to encode temporal information, has a better performance than the linear regression and
nonlinear regression models such as the artificial neural network model. Linear regression is a linear
approach to modeling the relationship among different sensors. The artificial neural network model,
which has great nonlinear curve fitting capability, can achieve a better performance than the linear
regression model. Due to the high correlation between sensors in the same group and the strong
feature extract ability of the LSTM model, either of the two sensors can achieve a better performance
on fault diagnosis. Virtual sensors are used to diagnose faults in this paper, which are constructed by
the linear regression (LR) model, the artificial neural network (ANN) model and the LSTM model.
Tsuc and Psuc with minimum biases are chosen as instances.
The performance of virtual sensors can be verified by the different grouping thresholds.
In Figure 7, the absolute deviation and fault diagnosis ratios obviously increase along with increase
of grouping threshold. The absolute deviation of the LSTM-based virtual sensor has significant
performance than FC-based and LR-based virtual sensor, when the fault samples occur. For instance,
the average absolute deviation between pressure sensor Psuc and virtual sensors will decrease from
0.027 to 0.021, if LSTM-based virtual sensor is replaced by FC-based and the grouping threshold takes
0.8. The average absolute deviation is decreased from 0.027 to 0.019 and the fault diagnosis ratios is
decreased from 100.0% to 95.0% if the LSTM-based virtual sensor is similarly replaced by LR-based.
When the fault samples occur, increasing absolution deviation can be obtained along with the increase
of the grouping thresholds. For instance, the average absolute deviation is decreased from 0.021 to
0.027, if the grouping threshold takes 0 instead of 0.8. It reveals that LSTM-based virtual sensors have
an excellent performance along with the increase of the grouping threshold.
As shown in Figure 8, average fault diagnosis ratio are compared in different biases. For these
two sensors, the fault diagnosis ratios of three methods increases with the increase of the introduced
biases fault in Table 4. All methods easily reach 100% when the magnitude of biases is the maximum of
each sensor. However, as the magnitude of biases decreases, the fault diagnosis rates show a difference.
For instance, when the biases of the temperature sensor Tsuc is tiny, such as 0.5, FC-based and LR-based
virtual sensors have a poor performance while the fault diagnosis ratio of the LSTM-based virtual
sensor reaches 98%. It means that LSTM-based virtual sensors have a better performance in the
case of the positive or negative biases. When the biases are negative, all methods are influenced to
various extents. For instance, average fault diagnosis ratios of the pressure sensor Psuc are 98.5%, 94%
and 89.75% respectively, provided that magnitude of biases is changed from 0.5 to −0.5. However,
fault diagnosis ratios of LSTM-based virtual sensors always achieve a better performance than
other methods.
To further verify the performance of another virtual sensor during training and testing, the mean
absolute error (MAE) is used in this paper. MAE is frequently used to measure the differences between
values predicted by a model and the values observed. The MAE between the physical sensor and
virtual sensor is defined as

N
1
MAE =
N ∑ (xi − x̂i ), (15)
i =1
Sensors 2019, 19, 3013 11 of 15

where N represents sampling steps, xi represents value of virtual sensor and x̂i represents value
of physical sensor. In actual implementation, the MAE of the corresponding sensor is obtained by
calculating the average of the two virtual sensors.

0.6 0.6

The absolute deviation

The absolute deviation

LSTM LSTM
ANN ANN
LR LR
0.4 0.4

0.2 0.2

0.0 0.0

Time series Time series

(a) (b)
0.04 0.04
The absolute deviation

LSTM

The absolute deviation

LSTM
ANN ANN
0.03 LR 0.03 LR

0.02 0.02

0.01 0.01

0.00 0.00

Time series Time series

(c) (d)

0.4 0.4
The absolute deviation
The absolute deviation

LSTM LSTM
ANN ANN
0.3 LR 0.3 LR

0.2 0.2

0.1 0.1

0.0 0.0

Time series Time series

(e) (f)
0.04 0.04
The absolute deviation

LSTM
The absolute deviation

LSTM
ANN ANN
0.03 LR 0.03 LR

0.02 0.02

0.01 0.01

0.00 0.00
Time series Time series
(g) (h)
0 when
Figure 7. The absolute deviation of Tsuc and Psuc . (a) The absolute deviation of virtual sensor Tsuc
00
the grouping threshold takes 0.8; (b) The absolute deviation of virtual sensor Tsuc when the grouping
threshold takes 0.8; (c) The absolute deviation of virtual sensor Psuc 0 when the grouping threshold
00
takes 0.8; (d) The absolute deviation of virtual sensor Psuc when the grouping threshold takes 0.8;
0 when the grouping threshold takes 0; (f) The absolute
(e) The absolute deviation of virtual sensor Tsuc
00
deviation of virtual sensor Tsuc when the grouping threshold takes 0; (g) The absolute deviation of
0 when the grouping threshold takes 0; (h) The absolute deviation of virtual sensor
virtual sensor Psuc
00
Psuc when the grouping threshold takes 0.
Sensors 2019, 19, 3013 12 of 15

100 100

Fault diagnosis ratios (%) 95

Fault diagnosis ratios (%)

95

90 90

85 85

80 80

75 75
LR LR
70 ANN ANN
LSTM 70 LSTM
0 0

-0.95 -0.8 -0.65 -0.5 0 0.5 0.65 0.8 0.95 -0.95 -0.8 -0.65 -0.5 0 0.5 0.65 0.8 0.95
（a） （b）

Figure 8. (a) Average fault diagnosis ratio of temperature sensor Tsuc with different biases. (b) Average
fault diagnosis ratio of temperature sensor Psuc with different biases.

As shown in Table 6, p-values between the traditional methods and the proposed method were
calculated. The t test can easily be adapted to testing the following hypotheses at a specified level of
significance α0 : H0 : µ1 = µ2 , H1 : µ1 6= µ2 . H0 will be rejected at any level α0 ≥ 0.05, which means
there is not a significant difference between the two methods. As we can see, p-values between
the traditional methods and the proposed method are always smaller than 0.05. This means that
LSTM-based virtual sensors are significantly different to the two other methods.

Table 6. P-values between the traditional methods and the proposed method.

Tdis Tcon_out Tre f _b Tre f _a Tchw_s Tchw_r Pdis Pth_in Pth_out

LR-based, LSTM-based 1.8 × 10−5 3.5 × 10−3 1.3 × 10−2 5.4 × 10−4 1.4 × 10−2 3.3 × 10−3 2.1 × 10−3 6.2 × 10−4 1.5 × 10−3
FC-based, LSTM-based 2.0 × 10−2 1.6 × 10−2 2.8 × 10−4 3.7 × 10−4 2.3 × 10−2 2.9 × 10−3 3.0 × 10−2 5.8 × 10−4 2.9 × 10−3

In addition to Tsuc and Psuc , Figure 9 shows MAE of another nine sensors during training when
the grouping threshold takes 0.8. Smaller value of MAE indicates better training performance of the
virtual sensors. As we can see, due to the different reading range of different sensors, virtual sensors
show different performances for different physical sensors. However, LSTM-based virtual sensors
always have a lower MAE value compared with LR-based or FC-based virtual sensors. It reveals that
LSTM-based virtual sensors always have a better training performance. Table 7 summarizes MAE
of another nine sensors during testing, to which minimum positive biases are added. The grouping
threshold also takes 0.8. On the contrary, the maximal value of MAE indicates better training
performance of the virtual sensors. When the minimum negative bias occurred, virtual sensors
based on different models show a different performance. Therefore, LSTM-based virtual sensors have
a better performance for both positive and negative biases.

Table 7. MAE of the remaining sensors during testing.

Tdis Tcon_out Tre f _b Tre f _a Tchw_s Tchw_r Psuc Pth_in Pth_out

LR-based 1.37 0.89 0.87 0.42 0.44 0.38 0.061 0.057 0.032
Positive FC-based 1.08 0.84 0.85 0.29 0.21 0.27 0.049 0.053 0.021
LSTM-based 0.85 0.68 0.69 0.17 0.19 0.20 0.034 0.038 0.016
LR-based 1.41 0.82 0.89 0.33 0.47 0.39 0.067 0.062 0.035
Negative FC-based 1.12 0.89 0.81 0.33 0.27 0.31 0.052 0.058 0.018
LSTM-based 0.87 0.56 0.75 0.20 0.21 0.25 0.032 0.040 0.019
Sensors 2019, 19, 3013 13 of 15

1.4
LSTM LSTM
ANN 0.06 ANN
1.2 LR LR

0.05
1.0

0.8 0.04

MAE
MAE

0.6
0.03

0.4
0.02
0.2
0.01
0.0
Tdis Tcon -out Tref -b Tref - a Tchw- s Tchw- r Pdis Pth -in Pth-oout
(a) (b)

Figure 9. (a) MAE of the remaining temperature sensors during training. (b) MAE of the remaining
pressure sensors during training.

In order to verify the performance of different physical sensors, fault diagnosis ratios of all
physical sensors with minimum biases are listed in Tables 8 and 9. Table 8 lists fault diagnosis
ratios when the grouping threshold takes 0.8 and the minimum positive biases are added. It can be
clearly seen that LSTM-based virtual sensors have a better performance than another two methods.
This conclusion is consistent with the results of Figure 9 and Table 7. For instance, fault diagnosis
rate with positive bias is 91.00% and 96.25% rather than 98.75%, when the LSTM-based virtual sensor
is similarly replaced by LR-based and FC-based virtual sensor for physical sensor Tcon_out . The fault
diagnosis ratios are reduced by 7.75% and 2.5% compared to LR-based and FC-based virtual sensor,
respectively. When the minimum negative biases occur, LSTM-based virtual sensors still have a better
performance in Table 9.

Table 8. Fault diagnosis ratios with positive bias.

Tsuc Tdis Tcon_out Tre f _b Tre f _a Tchw_s Tchw_r Psuc Pdis Pth_in Pth_out
LR-based 90.5% 92.25% 91.0% 94.5% 93.25% 89.5% 90.5% 92.5% 93.0% 95.25% 94.5%
FC-based 93.75% 93.5% 96.25% 95.75% 95.5% 91.75% 97.5% 96.0% 95.5% 96.25% 97.75%
LSTM-based 98% 98.5% 98.75% 97.25% 97.75% 96.25% 100.0% 97.25% 99.25% 97.5% 99.5%

Table 9. Fault diagnosis ratios with negative bias.

Tsuc Tdis Tcon_out Tre f _b Tre f _a Tchw_s Tchw_r Psuc Pdis Pth_in Pth_out
LR-based 90.25% 92.5% 89.25% 94.25% 92.5% 86.5% 86.25% 89.75% 94.5% 95.0% 94.25%
FC-based 91.25% 96.75% 92.25% 96.0% 95.5% 92.25% 93.5% 94.0% 96.75% 97.5% 96.75%
LSTM-based 95.25% 98.75% 93.0% 95.5% 97.5% 96.75% 95.25% 98.5% 98.5% 99.25% 97.25%

5. Conclusions
The operating efficiency of an air-cooled chiller system is critical for building energy performance.
In this paper, a novel chiller sensor fault diagnosis method was proposed. Chiller sensors were
divided into different groups by MIC score and the grouping threshold. Highly correlated sensors can
significantly improve the accuracy of virtual sensors. Under the same model, experimental results
also show that virtual sensors have a better performance by grouping with the MIC score. Virtual
Sensors 2019, 19, 3013 14 of 15

sensors constructed by LSTM model were compared with the linear regression and the artificial neural
network model, which reduces the cost of hardware redundancy. In order to reduce the false alarm
rate and directly diagnose sensor faults, two virtual sensors corresponding to the same physical sensor
were constructed. Comparing the method in this paper with using only one virtual sensor, the method
in this paper has the lower false alarm rate. The proposed method with higher accuracy has also been
well presented and experimentally validated.
The following work needs to be done in the future: firstly, fault diagnosis threshold determination
is an indispensable part of the chiller sensor fault diagnosis based on deviation. Threshold will
directly impact the performance of fault diagnosis. An excellent method for fault diagnosis threshold
should be further researched; secondly, the proposed method will be applied to other refrigeration air
conditioning systems to further verify generalization.

Author Contributions: Conceptualization, L.G. and D.L. (Donghui Li); methodology, L.G.; software, L.G.;
validation, L.G., D.L. (Ding Li) and L.L.; formal analysis, L.G., Y.G. and D.L. (Ding Li); investigation, L.G. and
L.L.; resources, L.G. and L.L.; data curation, L.G.; writing—original draft preparation, L.G.; writing—review and
editing, L.G., D.L. (Ding Li) and L.Y.; visualization, L.G.; supervision, D.L. (Donghui Li); project administration,
L.G.; funding acquisition, D.L. (Donghui Li)
Funding: This work is supported by the National Natural Science Foundation of China under Grant 61873180.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Brambley, M. Review article: Methods for fault detection, diagnostics, and prognostics for building
systems—A review, Part II. HVAC R Res. 2005, 11, 19.
2. Zheng, Y.; Ghahramani, A.; Becerik-Gerber, B. Building occupancy diversity and HVAC (heating, ventilation,
and air conditioning) system energy efficiency. Energy 2016, 109, 641–649.
3. Saidur, R.; Hasanuzzaman, M.; Mahlia, T.M.I.; Rahim, N.A.; Mohammed, H.A. Chillers energy consumption,
energy savings and emission analysis in an institutional buildings. Energy 2011, 36, 5233–5238.
4. Katipamula, S.; Brambley, M. Review article: Methods for fault detection, diagnostics, and prognostics for
building systems—A review, Part I. Hvac R Res. 2005, 11, 3–25.
5. Li, Z.; Li, Q.; Wu, Z.; Jie, Y.; Zheng, R. A fault diagnosis method for on load tap changer of aerospace power
grid based on the current detection. IEEE Access 2018, 6, 1.
6. Ma, J.; Jin, J. Applications of fault detection and diagnosis methods in nuclear power plants: A review.
Prog. Nucl. Energy 2011, 53, 255–266.
7. Chi, M.V.; Wong, P.K.; Wong, K.I. Simultaneous-fault detection based on qualitative symptom descriptions
for automotive engine diagnosis. Appl. Soft Comput. 2014, 22, 238–248.
8. Dinca, L.; Aldemir, T.; Rizzoni, G. A model-based probabilistic approach for fault detection and identification
with application to the diagnosis of automotive engines. IEEE Trans. Autom. Control 2002, 44, 2200–2205.
9. Kim, W.; Katipamula, S. A review of fault detection and diagnostics methods for building systems.
Sci. Technol. Built Environ. 2017, 24, 3–21.
10. Gao, Z.; Cecati, C.; Ding, S.X. A survey of fault diagnosis and fault-tolerant techniques—Part I: Fault
diagnosis with model-based and signal-based approaches. IEEE Trans. Ind. Electron. 2015, 62, 3757–3767.
11. Gao, Z.; Cecati, C.; Ding, S.X. A survey of fault diagnosis and fault-tolerant techniques—Part II: Fault
diagnosis with knowledge-based and hybrid/active approaches. IEEE Trans. Ind. Electron. 2015,
62, 3768–3774.
12. Ren, N. PCA-SVM-based automated fault detection and diagnosis (AFDD) for vapor-compression
refrigeration systems. Hvac R Res. 2010, 16, 295–313.
13. Li, D.; Li, D.; Li, C.; Li, L.; Gao, L. A novel data-temporal attention network based strategy for fault diagnosis
of chiller sensors. Energy Build. 2019, 198, 377–394, doi:10.1016/j.enbuild.2019.06.034.
14. Bo, F.; Du, Z.; Jin, X.; Yang, X.; Guo, Y. A hybrid FDD strategy for local system of AHU based on artificial
neural network and wavelet analysis. Build. Environ. 2010, 45, 2698–2708.
15. Benmoussa, S.; Djeziri, M.A. Remaining useful life estimation without needing for prior knowledge of the
degradation features. IET Sci. Meas. Technol. 2017, 11, 1071–1078, doi:10.1049/iet-smt.2017.0005.
Sensors 2019, 19, 3013 15 of 15

16. Ms, L.H.C.; Russell, E.L.; Braatz, R.D. Fault Detection and Diagnosis in Industrial Systems; Springer:
London, UK, 2001.
17. Yu, Y.; Woradechjumroen, D.; Yu, D. A review of fault detection and diagnosis methodologies on air-handling
units. Energy Build. 2014, 82, 550–562.
18. Mattera, C.G.; Quevedo, J.; Escobet, T.; Shaker, H.R.; Jradi, M. A method for fault detection and diagnostics
in ventilation units using virtual sensors. Sensors 2018, 18, 3931.
19. Ergan, S.; Radwan, A.; Zou, Z.; Tseng, H.A.; Han, X. Quantifying human experience in architectural spaces
with integrated virtual reality and body sensor networks. J. Comput. Civ. Eng. 2018, 33, 04018062.
20. Tegen, A.; Davidsson, P.; Mihailescu, R.C.; Persson, J. Collaborative sensing with interactive learning using
dynamic intelligent virtual sensors. Sensors 2019, 19, 477.
21. Wei, C.; Chen, J.; Song, Z.; Chen, C.I. Adaptive virtual sensors using SNPER for the localized construction
and elastic net regularization in nonlinear processes. Control Eng. Pract. 2019, 83, 129–140.
22. Valdivia, C.G.; Escobedo, J.L.C.; Durán-Muñoz, H.A.; Berumen, J.; Hernandez-Ortiz, M. Implementation of
virtual sensors for monitoring temperature in greenhouses using CFD and control. Sensors 2019, 19, 60.
23. Yoon, S.; Yu, Y. Strategies for virtual in-situ sensor calibration in building energy systems. Energy Build.
2018, 172, 22–34.
24. Li, H.; Yu, D.; Braun, J. A review of virtual sensing technology and application in building systems.
Hvac R Res. 2011, 17, 619–645.
25. Mattera, C.G.; Quevedo, J.; Escobet, T.; Shaker, H.R.; Jradi, M. Fault detection and diagnostics in ventilation
units using linear regression virtual sensors In Proceedings of the International Symposium on Advanced
Electrical and Communication Technologies (ISAECT), Rabat, Morocco, 21–23 November 2018.
26. Reppa, V.; Papadopoulos, P.; Polycarpou, M.M.; Panayiotou, C.G. A distributed virtual sensor scheme for
smart buildings based on adaptive approximation. In Proceedings of the 2014 International Joint Conference
on Neural Networks (IJCNN), Beijing, China, 6–11 July 2014; pp. 99–106.
27. Kusiak, A.; Li, M.; Zheng, H. Virtual models of indoor-air-quality sensors. Appl. Energy 2010, 87, 2087–2094.
28. Kanjo, E.; Younis, E.; Ang, C.S. Deep learning analysis of mobile physiological, environmental and location
sensor data for emotion detection. Inf. Fusion 2018, 49, 46–56.
29. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the 2016
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June
2016; pp. 770–778.
30. Shin, H.; Roth, H.R.; Gao, M.; Lu, L.; Xu, Z.; Nogues, I.; Yao, J.; Mollura, D.; Summers,
R.M. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset
characteristics and transfer learning. IEEE Trans. Med. Imaging 2016, 35, 1285–1298.
31. Vinh, N.X.; Epps, J.; Bailey, J. Information theoretic measures for clusterings comparison: Variants, properties,
normalization and correction for chance. J. Mach. Learn. Res. 2010, 11, 2837–2854.
32. Reshef, D.N.; Reshef, Y.A.; Finucane, H.K.; Grossman, S.R.; Gilean, M.V.; Turnbaugh, P.J.; Lander, E.S.;
Michael, M.; Sabeti, P.C. Detecting novel associations in large data sets. Science 2011, 334, 1518–1524.

c 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).